Integrated Data
Analysis and Visualization
Group 5 Report
DOE Data Management Workshop
May 24-26, 2004
Chicago, IL
Group 5 Participants
Wes Bethel, LBNL
George Michaels, PNNL
John Blondin, NC State
Habib Najm, SNL
Julian Bunn, Caltech
John van Rosendale, DOE/HQ
George Chin, PNNL
Nagiza Samatova, ORNL
Chris Ding, LBNL
David Schissel, General Atomics
Irwin Gaines, FNAL
Gary Strand, NCAR
Chandrika Kamath, LLNL
Todd Smith, Geospiza, Inc.
James Kohl, ORNL
Outline
 What is integrated data analysis and
visualization? Why do we care?
 Data Complexity & Implications
 Applications-driven capabilities
 Technology gaps to address these capabilities
 General Recommendations
 Conclusions
The Curse of Ultrascale Computation
and High-throughput Experimentation
Computational and experimental advances enable capturing of complex
natural phenomena, on a scale not possible just a few years ago. With this
opportunity, comes a new problem – the petabyte quantities of produced
data. As a result, answers to fundamental questions about the nature of the
universe largely remain hidden in these data.
How to enable scientists perform analyses and
visualizations of these raw data to extract knowledge?
Tony’s Scenario
Data Select  Data Access  Correlate  Render  Display
(density, pressure)
From astro-data
Where (step=101)
(x-velocity>Y);
Sample (density, pressure)
Scientific Process
Automation Layer
Data Mining &
Analysis Layer
Storage Efficient
Access Layer
Run
analysis
Run viz filter
Visualize
scatter plot
Workflow Design & Execution
Select
Data
Take
Sample
Use Bitmap Get variables
(condition) (var-names, ranges)
Bitmap
Index
Selection
Analysis
Tool
Read Data
(buffer-name)
Write Data
VIZ
Tool
Read Data
Read Data
(buffer-name) (buffer-name)
Write Data
Parallel
HDF
Hardware, OS, and MSS (HPSS)
PVFS
Integration Must Happen at
Multiple Levels
 To enable end-to-end system
performance, 80-20 rule and novel
discoveries integration must happen:

Between and within data flow levels:

Workflows  Analysis & Viz  Access &
Movement)

Across geographically distributed resources

Across multiple data scales and resolutions
Challenge of Data Massiveness
Drinking from the firehose


Climate

Now: 20-40TB per simulated year

5 yrs: 100TB/yr 5-10PB/yr
Fusion




Now: 100Mbytes/15min
5 yrs: 1000Mbytes/2 min with realtime
comparison with running experiment,
500Mbits/sec guaranteed (QoS)
High Energy Physics

Now: 1-10PB data stored, Gigabit net.

5 yrs: 100PB data, 100Gbits/sec net
Chemistry (Combustion and Nanostructures)

Now: 10-30TB data

Sharf’smulticast
stats, LBL)
5 yrs: 30-100TB data, (John
10Gbits/sec
Most of this Data will NEVER Be Touched
with the current trends in technology
 The amount of data stored online quadruples every 18
months, while processing power ‘only’ doubles every 18
months.

Unless the number of processors increases unrealistically rapidly,
most of this data will never be touched.
 Storage device capacity doubles every 9 months, while
memory capacity doubles every 18 months (Moore’s law).

Even if the divergence between these rates of growth will
converge, the memory latency is and will remain the rate-limiting
step in data-intensive computations
 Operating systems struggle to handle files larger than a few
GBs.

OS constraints and memory capacity determine data set file size
and fragmentation
Challenge of Breaking the
Algorithmic Complexity Bottleneck
MS Data Rates:
100’sGB10’sTB/day(2004)1.0’sPB/day(2008)
Algorithm Complexity
Algorithmic Complexity:
Calculate means
Calculate FFT
Data
size, n
n
nlog(n)
n2
100B
10-10sec.
10-10 sec.
10-8 sec.
10KB
10-8 sec.
10-8 sec.
10-4sec.
1MB
10-6 sec.
10-5 sec.
1 sec.
100MB
10-4 sec.
10-3 sec.
3 hrs
10GB
10-2 sec.
0.1 sec.
3 yrs.
O(n)
O(n log(n))
Calculate SVD
O(r • c)
Clustering algorithms
O(n2)
For illustration chart assumes 10-12 sec.
calculation time per data point
Massive Data Sets are Naturally
Distributed BUT Effectively Immoveable
(Skillicorn, 2001)

Bandwidth is increasing but not at the same rate as stored data


Latency for transmission at global distances is significant


There are some parts of the world with high available bandwidth BUT there are
enough bottlenecks that high effective bandwidth is unachievable across
heterogeneous networks
Most of this latency is time-of-flight and so will not be reduced by technology
Data has a property similar to inertia:

It is cheap to store and cheap to keep moving, but the transitions between
these two states are expensive in time and hardware.

Legal and political restrictions

Social restrictions

Data owners may let access data but only by retaining control of it
• Should we move computations to the data, rather than data to the
computations?
• Should we cache the data close to analysis and viz.?
• Should we be smarter about reducing the size of the data while
having the same or richer information content?
Challenge of High Dimensionality,
Multi-Scale and Multi-Resolution
Time
The experiment
paradigm is changing to
statistically capture the
complexity. We will get
maximum value when we
explore three or more
dimensions/scales in a
single experiment.
(From G. Michaels, PNNL)
But multi-scale and multi-resolution analysis and
visualization are in their infancy!
Know Our Limits & Be Smart
Obligations are two-sided: CS and Apps
Not humanly possible to browse a petabyte of data.
Analysis must select views or reduce to quantities of interest
rather than push more views past the user.
More data
Ultrascale Simulations:
Must be smart about which probe
combinations to see!
Physical Experiments:
Must be smart about probe placement!
Petabytes
Can we browse a petabyte of data?
Terabytes
To see 1 percent of a petabyte at 10
megabytes per second takes 35 8-hour
days!
Gigabytes
Megabytes
No
Analysis
Region
Selection
Analysis-driven
summarization
More
analysis
Analysis of full context must select views
or reduce to quantities of interest in
addition to fast rendering of data.
Frame of Context Differences Suggest
Needs in Hardware and Software for
Analysis and Visualization
Arguably, visualization can be the most critical step of a
simulation experiment. But it has to be in a full context.
Visualization
I hear and I forget.
I see and I believe.
I do Visual Analysis and I understand.
—Confucius (551-479 BC)
Frames of context for major
steps of space-time simulation
scientific discovery process.
Time
Space
Simulation
Analysis
Storage
Need hardware and software for
Full Context analysis and
visualization
(From G. Ostrouchov, ORNL)
But Tony Still Has a Dream –
Internet “Plug-ins” for Ultrascale Computing!
Paraview
ASPECT
From Dreams to
Achievable Applicationdriven Capabilities
The first step in Group 5 discussion
Representation by Applications:
Technology Gaps Ranking:
• Climate
• Biology
• Combustion
• Fusion
• HENP
• Supernova
Research & Development -- 3
Hardening Technology -- 2
Deployment & Maintenance -- 1
Capability #1: IDL-like SCALABLE,
open source environment (J.Blondin)

High performance-enabling technologies:

Parallel analysis and viz. algorithms (e.g. pVTK, pMatlab, parallel-R) (3—2)

Portable implementation on HPC platforms (2—1)

Hardware accelerated implementations (GPUs, FPGAs) (2—1)

Parallel I/O libraries coupled with analysis and viz (ROMIO+pVTK, pNetCDF+Parallel-R)
(2—1)

Information visualization (3—2)

Interoperability-enabling technologies:


Component architectures (CCA) (3—2)

Core data models and data structures unification (3—2)

Structural, semantic and syntactic mediation (3—2)
Scripting environments:

IDL/Matlab-like high-level programming languages (3)

Optimized (parallel, accelerated) functions (core libraries) (3)

Simulation interfaces (3)

Visualization interfaces (information, statistical and scientific visualization) (3—2)
Capability #2: Domain-specific
libraries and tools
 Same technologies as for Capability #1 Plus
More
 Novel algorithms for domain-specific analysis
and visualization:

Feature extraction/selection/tracking (e.g., ICA for
climate) (3—2)

New types of data (e.g. trees, networks) (3—2)

Interpolation & transformation (3—2)

Multi-scale/hierarchical features correlation (3)
 Novel data models if necessary (3)
Capability #3: “Plug and Play”
Analysis and Visualization Envs
 Community-specific data model(s) (3)

Standardization that still provides efficiency
and flexibility
 Community-specific common APIs (3)

Unified still extensible data structures
 Common component architectures (3—2)
“i”-ntegration vs. “I”-ntegration
Components integration strategies should be assessed
within single and between multiple higher-level applications


“i”-integration within the same application:

Same set of data structures => brute-force check (at worst)

Same language, execution and control model

Scripting languages (TCL, Python, R)

Run on the same cluster of machines
“I”-integration across multiple applications:

Different data structures (unknown for future apps)

Different execution & control models

Different programming languages

Run on different hosts
Data Formats
Transformations:
•
•
•
•
File  App
App_X  App_X
App_X  App_Y
App  File
Capability #4: Feature (region)
detection, extraction, tracking
 Efficient and effective data indexing that
(3-2):



Supports unstructured data in files (e.g.,
bitmap indexing extension to AMR)
Supports heterogeneous/non-scalar data (e.g.,
vector fields, protein sequence, protein
function, pathway, network)
Supports on-demand derived data (e.g.
F(X)/G(Y)<5: entropy as a function of indexed
density and pressure)
 Information visualization (3—2)
Other Capabilities

Remote, collaborative & interactive analysis and visualization (3-2):

Network-aware analysis and viz

Novel means of hiding latency (e.g. caching via LoCi, view-dependent
isosurfaces)

Sensitivity & uncertainty quantification (3)

Streaming analysis & viz (3):


Approx. multi-res. algorithms

Data transformations on streaming data
Annotation & provenance of analysis and visualization results

System-, analysis & viz-, data-level metadata

Verification and validation

Comparative analysis and visualization

Cross-cutting capabilities:

Integration of analysis and visualization with workflows

Integration of analysis and visualization with data bases (e.g., query-based)
General Recommendations
 Encourage open source software
 Move out mature technologies (??)
 Encourage/force data model(s) & APIs
standardization efforts?
 Do not expect scientists to develop their
domain-specific components rather fund
collaborative CS & Apps teams  Will assure
more robust and reusable solutions and take
the burden of CS tasks from scientists
Conclusions
 Integration must occur at multiple levels
 Integration is more easily achievable
within a community than across
communities
 Community-based data model(s) and
APIs are required for “Plug & Play”